v0.4.0-alpha.9

Requires DeepSeek Harness >=0.1.2-alpha.3 <0.2.0. Experimental local ASR supports Apple Silicon Macs and unsigned Windows x64. Mixed rc.2/Alpha.2 profile dependencies may report peer warnings; use an Alpha.3-compatible profile.

What's Changed

• docs: align Alpha.8 release state and gate installs by @franksong2702 in #17
• feat: add optional Composer hold-to-talk by @franksong2702 in #18
• release: prepare dsh-dictate 0.4.0-alpha.9 by @franksong2702 in #19

Full Changelog: v0.4.0-alpha.8...v0.4.0-alpha.9

v0.4.0-alpha.8

Requires DeepSeek Harness >=0.1.2-alpha.3 <0.2.0. Experimental local ASR supports Apple Silicon Macs and unsigned Windows x64. Mixed rc.2/Alpha.2 profile dependencies may report peer warnings; use an Alpha.3-compatible profile.

What's Changed

• fix: make alpha release recovery digest-safe by @franksong2702 in #13
• feat: add experimental Windows x64 local ASR for DSH Alpha.3 by @franksong2702 in #8
• docs: update Windows and DSH Alpha.3 guidance by @franksong2702 in #14
• release: prepare dsh-dictate 0.4.0-alpha.8 by @franksong2702 in #15
• fix: treat packed release artifact as a local file by @franksong2702 in #16

Full Changelog: v0.4.0-alpha.7...v0.4.0-alpha.8

v0.4.0-alpha.7

Requires DeepSeek Harness >=0.1.2-alpha.2 <0.2.0. Local ASR currently supports Apple Silicon Macs.

What's Changed

• v0.4.0-alpha.7 compat: support official DSH 0.1.2-alpha.2 by @franksong2702 in #11

Full Changelog: v0.4.0-alpha.6...v0.4.0-alpha.7

dsh-dictate v0.4.0-alpha.6

DSH Dictate 0.4.0-alpha.6

Requires DeepSeek Harness >=0.1.1-rc.2 <0.2.0. Managed local ASR supports Apple Silicon Macs; Windows local ASR is not included in this release.

Changes

• One opaque Composer focus card shows the current stage and non-final transcript preview. The upper history stack has been removed.
• Long previews scroll without moving the conversation. Recording and processing use breathing indicators; successful completion briefly stays visible, then slides down into the Composer.
• Web Speech and local ASR share a nine-minute recording limit with a warning countdown from eight minutes. Reaching the limit ends recording and processes the result without automatically sending it.
• Local ASR audio is transcribed in bounded segments. Web Speech reconnects after browser-ended sessions and handles unsupported phrase hints more gracefully.
• GitHub Actions publishes the alpha through npm Trusted Publishing with provenance.

Known issues and validation boundary

During a recent speaker-to-microphone demo on the test setup, local transcription returned an incomplete phrase and model polishing failed with the selected Qwen3.5 9B / LM Studio route. The cause is still under investigation; that attempt did not pass end-to-end validation. Review the resulting draft before sending and keep automatic sending disabled if unsure.

The README and demonstration GIF refresh are deferred and are not part of this release.

中文说明

要求 DeepSeek Harness >=0.1.1-rc.2 <0.2.0。本地 ASR 一键安装目前支持 Apple Silicon Mac，本次不包含 Windows 本地 ASR。

• Composer 上方统一为一个不透明焦点卡片，只显示当前阶段及非最终转写预览，移除上方历史信息。
• 长文本可滚动，不推动聊天记录；录音和处理阶段显示呼吸指示灯，成功完成后短暂停留，再向下收起。
• 两种识别方式统一设置九分钟录音上限，从八分钟开始倒计时提醒。到时自动结束并处理，但不会自动发送。
• 本地 ASR 使用分段转写；改善浏览器语音识别的自动续接及不支持词汇提示时的处理。
• 通过 GitHub Actions 与 npm Trusted Publishing 发布，并提供来源证明。

已知问题：最近一次测试环境的扬声器转麦克风演示出现了转写不完整，以及所选 Qwen3.5 9B / LM Studio 路线润色失败，原因尚未定位。这次演示不算端到端验证通过。请在发送前检查输入框中的结果，不确定时保持自动发送关闭。

README 和演示 GIF 本次不更新，后续补充。

Full changelog: v0.4.0-alpha.5...v0.4.0-alpha.6

dsh-dictate v0.4.0-alpha.5

What changed

• Adds a one-click local SenseVoice setup for Apple Silicon Macs. The native runtime ships in the plugin package; users no longer need Python, PATH changes, an internal install source, or a visible service address.
• Downloads the pinned SenseVoice Q8 model from the settings page with resume support, SHA-256 verification, startup, prewarm, status checks, and opt-in “start with DSH”.
• Reuses an already verified model when the bundled runtime is upgraded, while refusing to start a runtime that does not match the current package.
• Keeps browser Web Speech as the zero-setup default and makes browser/local recognition choices explain installation, network, and audio-processing differences before selection.
• Removes the persistent automatic-fallback setting. If local recognition is unavailable before recording, switching to browser recognition now requires an explicit one-time action.
• Groups voice settings into recognition, input, context enhancement, and sending sections, with an updated settings screenshot.
• Rejects browser requests from origins other than the active loopback DSH instance before accepting audio.

Compatibility

• Requires DeepSeek Harness 0.1.1-rc.2.
• One-click local ASR currently supports macOS Apple Silicon only. macOS Intel, Windows, and Linux continue to use Web Speech.
• Local recognition is stop-then-transcribe in this alpha; streaming partial transcription remains out of scope.

Verification

• Typecheck and full test suite: 6 files, 111 tests passed.
• Native runtime: 2 Rust tests passed; Clippy with warnings denied; formatting and ad-hoc signature verified.
• Package verification: dsh-dictate@0.4.0-alpha.5, 42 package entries, bundled runtime SHA-256 dab83ea0c5bfa95b8e9c94f804da7d88c9fc5657ac5ac8503554ec338d5db52f.
• Real installation and microphone acceptance on DSH port 3081; final endpoint transcription returned 这是一次本地语音识别测试.
• Trusted DSH origin returned HTTP 200; an untrusted browser origin returned HTTP 403.

Alpha limitation

The bundled Apple Silicon runtime is ad-hoc signed and is not Apple Developer ID notarized.

This is an alpha prerelease. Community project; not affiliated with or endorsed by DeepSeek or DeepSeek Harness.

dsh-dictate v0.4.0-alpha.4

What changed

• Keeps browser Web Speech as the default transcription route and preserves the optional local SenseVoice endpoint route.
• Updates the plugin peer and development dependencies to DeepSeek Harness 0.1.1-rc.2.
• Adds explicit local-service lifecycle handling for 127.0.0.1:39081, including concurrent-start serialization and CORS-origin reconciliation.
• Adds settings-page service/install status, cancellation, retry, and opt-in “start with DSH” controls.
• Keeps the native macOS arm64 installer as an internal test entry; public users can prepare and manage their own funasr-server endpoint.

Compatibility

• Requires DeepSeek Harness 0.1.1-rc.2.
• Web Speech remains the zero-setup default.
• Local SenseVoice uses a loopback endpoint and does not run the model in the browser.

Verification

• Typecheck and full test suite: 6 files, 110 tests passed.
• Package verification: dsh-dictate@0.4.0-alpha.4, 40 package entries, host exports and client wrapper verified.
• Real managed-install E2E on 127.0.0.1:39081: HTTP 200, transcript 开饭时间早上九点至下午五点.

This is an alpha prerelease. Community project; not affiliated with or endorsed by DeepSeek or DeepSeek Harness.

dsh-dictate v0.4.0-alpha.3

What changed

• Adds an optional loopback-only local SenseVoice transcription path while keeping browser Web Speech as the default.
• Adds plugin-settings controls for starting, stopping, inspecting, and testing the fixed 127.0.0.1:39081 service.
• Adds voice-activity feedback, about 600 ms of ending-audio protection, clear Composer stage labels, and explicit fallback policies.
• Keeps already-recorded audio from being silently rerouted to another service.
• Refreshes the README, settings screenshot, package metadata, and bilingual promotion copy.

Local mode boundary

Local mode is experimental and requires a local FunASR/SenseVoice environment. The first model download is close to 1 GB. Plugin-managed start, stop, and status controls apply only to 127.0.0.1:39081; custom loopback endpoints remain externally managed.

Install

dsh plugin --profile web add ./dsh-dictate-0.4.0-alpha.3.tgz

Restart the corresponding dsh web process after installation or upgrade.

Compatibility and verification

• Requires DeepSeek Harness 0.1.0-rc.7
• 5 test files, 98 tests passed
• Typecheck, package verification, and CI passed on Node 22.19 and Node 24
• Tarball installed without peer warnings into a fresh isolated DSH RC7 profile
• Isolated RC7 Web runtime exposed the dsh-dictate client entry

SHA-256: 18359321f026b5638d6fc389f6abf58f84367c735f403c902ad8e06d6f8a64e4

dsh-dictate v0.4.0-alpha.2

What changed

• Adds compatibility with DeepSeek Harness 0.1.0-rc.7.
• Migrates the plugin settings card to the RC7 namespace-key contract.
• Registers the matching Host settings section while keeping dictation preferences browser-local.
• Updates DSH package compatibility metadata to RC7.

Scope

This release keeps the existing browser Web Speech workflow, Session-aware phrase hints, optional model polishing, and user-controlled sending. Experimental local ASR and sherpa-onnx code are not included.

Install

dsh plugin --profile web add ./dsh-dictate-0.4.0-alpha.2.tgz

Restart the corresponding dsh web process after installation or upgrade.

Verification

• 3 test files, 59 tests passed
• typecheck, build, package verification, and pack passed
• CI passed on Node 22.19 and Node 24
• tarball installed without peer warnings into a fresh isolated DSH RC7 profile
• isolated RC7 Web runtime boot manifest included the dsh-dictate client entry

dsh-dictate v0.4.0-alpha.1

首个 dsh-dictate Alpha 预发布。

• 在 Composer 中实时显示语音转写
• 根据当前 Session 和 Composer 提取上下文词汇
• 支持可选模型润色和 Composer 录音快捷键
• 自动发送为 Beta 功能，默认关闭

安装：

dsh plugin --profile web add ./dsh-dictate-0.4.0-alpha.1.tgz

需要 Chrome 或 Edge，并面向 DeepSeek Harness 0.1.0-rc.6。完整功能、数据范围和限制见 README。

dsh-voice-input v0.3.0-alpha.9

第二个公开 Alpha 预发布。

新增

• 在 DSH Composer 上方显示实时最终文字与临时文字，停止后只写入 Composer 一次。
• 在“设置 → 插件 → 插件配置”中配置识别语言、模型润色和自动发送。
• 可选择当前 DSH 可用模型，使用最近 6 条可见用户/Assistant 文本进行 Session 上下文润色；上下文最多 12 KB。
• 可明确开启自动发送。只有用户点击结束录音才获得自动发送授权；浏览器自行结束识别时只填入 Composer。
• 转写完成状态在 3 秒后消失，发送后不再重复显示正文预览。

默认与隐私

• 模型润色与自动发送均默认关闭。
• 语音识别使用 Chrome 或 Edge Web Speech API；音频由浏览器语音服务处理，不经过 DSH Host。
• 开启模型润色后，原始转写和受限的可见 Session 文本会发送给所选模型提供商；不会发送系统提示词、工具、图片或 Assistant 推理。

已知限制

• 当前依赖 Chrome 或 Edge 的 Web Speech API。
• 模型润色辅助调用尚未写入 DSH Session 日志；上游开放相应扩展点后应迁移。
• 本版本通过 GitHub tarball 分发，不发布 npm。

安装

下载 dsh-voice-input-0.3.0-alpha.9.tgz，安装到 DSH Web profile 后重启对应的 dsh web 进程。

SHA-256：36e4da9bcb9c1f0e9b2d678e5f038bea52fce93a2ea622cccdd1c33865967fdd